89
CHAPTER 3
SYSTEMATIC ASSESSMENT
ON EXISTING WORKS
An efficient medical image retrieval system involves several stages
including preprocessing, feature extraction, classification, retrieval and
indexing. For retrieval of images, Euclidian distance method values are
calculated between the query image and database images. In all the existing
methods, various visual features have been considered indirectly to retrieve
the images from databases. This chapter presents implementation and result
analysis of some of the existing methods that are presented in literature
review in order to assess the exact retrieval performance of these works and
derive best out of them for further investigation.
Experimental setup: Intel® Core 2 Duo CPU Workstation with
2GB RAM used for conducting the experiments. MATLAB 7.2.0-Image
Processing tool Box used for developing user interface components as front
end, MATLAB 7.2.0-Image Processing tool Box-Workspace used as feature
database for backend storage and for image processing work other MATLAB
7.2.0 utilities were used. For mathematical equations, math type tool is used
for writing the document. Initially, MATLAB 7.2.0 workspace database with
several real time medical images were used for testing these CBMIR systems.
Similarity Comparison: In these works, for image retrieval the
similarity comparison technique has been computed. For similarity
comparison, Euclidean distance d has been computed using the following
equation
90
q db N 2(F [i] F [i])i 1
d (3.1)
where Fq[i] is the ith query image feature and Fdb[i] is the
corresponding feature in the feature vector database. Here N refers to the
number of images in the database (S.Nandagopalan et al 2008).
Retrieval Efficiency: Precision and recall are the basic measures
used in evaluating the search strategies of various methods. Precision
(accuracy of the retrieval) is the fraction of retrieved instances that are
relevant, while recall (sensitivity of the retrieval) is the fraction of relevant
instances that are retrieved.
Reference database: Precision and recall were calculated from the
reference database having 560 medical images that includes real-time medical
images collected from local hospital and medical images collected from
CasImage database and IRMA database repositories.
Analysis and Implementation: By analyzing the similarity in
execution and presentation concepts in the existing methods, it is being
concluded that all these works can be grouped into three categories viz.
single, dual and tri feature extraction.
For effective comparison of all the above there categories, they are
implemented in three parts under same experimental setup using same
reference database as follows:
1. CBMIR system using single visual feature content
2. CBMIR system using dual visual feature content
3. CBMIR system using tri visual feature content
91
The diagrammatic representation of these works is depicted in
Figure 3.1.
Figure 3.1 Framework for Systematic Assessment on Existing Works
3.1 CBMIR SYSTEM USING SINGLE VISUAL FEATURE
CONTENT
In this part, CBMIR system has been implemented using single
visual feature (either shape or texture) along with k-means clustering
algorithm. Section 3.1.1 discusses the implementation using shape based
feature and section 3.1.2 discusses it using texture based feature.
3.1.1 Shape based CBMIR system
In this work, basic shape feature is extracted using canny edge
detection algorithm. For classification, K-means classification algorithm is
used. The Framework model of the Shape based CBMIR system is shown in
Figure 3.2. Initially, medical images are taken as input to the system and
preprocessing of the images carried out in order to improve the flexibility of
the images for further processing of the system. In this system, there are two
CBMIR
Single Feature
Extraction
Dual Feature
Extraction
Tri Feature Extraction
Shape (CED)
Texture (GLCM)
Shape & Texture (EH& EBT)
Shape & Texture (GFD&GF)
Shape,Texture & Grayscale Resolution
(IM,GLCM & Grayscale
Resolution)
Shape,Texture & Intensity
(Gaussion Filter, LBP & DWT
Based Intensity Extraction)
92
parts of processes that are carried out namely offline and online. In offline
process, image reference database is constructed through one of the dominant
feature of the image such as shape extraction using Canny Edge Detection
(CED) algorithm. In online process, Graphical User Interface (GUI) for the
user interaction is developed through which the user can interact with the
system for retrieval of their desired images from image database. For retrieval
process, similarity comparison technique carried out between online user
query image and offline image reference database. After comparison, the
resulting images are indexed and retrieved based on their rank.
Shape
Shape is an important and most powerful feature used for image
classification, indexing and retrieval. Shape information is extracted using
histogram of edge detection. In this work, edge information in the image
obtained using canny edge detection (Canny, J 1986). Other techniques for
shape feature extraction are elementary descriptor, Fourier descriptor,
template matching, Quantized descriptors etc. Canny edge detection
outperforms many of the newer algorithms that have been developed in the
industry (Ehsan Nadernejad et al 2008).
93
Figure 3.2 Framework model of the Shape based CBMIR system
Algorithm: For detecting edges using Canny Edge Detection
Algorithm.
Step 1: Smoothing: Smooth the image with a two dimensional
Gaussian. In most cases the computation of a two dimensional Gaussian is
costly, so it is approximated by two one dimensional Gaussians.
Step 2: Finding Gradients: Take gradient of the image that shows
changes in intensity indicating the presence of edges. It actually gives two
results: the gradient in x direction and the gradient in y direction.
Step 3: Non-maximal suppression: Edges will occur at points
where the gradient is at a maximum. The magnitude and direction of the
gradient is computed at each pixel.
Input Medical Images
Image Enhancement
Reference Database
Query Image
Shape Extraction (CED)
Building Feature Vector
Extract Shape
Similarity Comparison
Indexing and Retrieval
Output Ranked Images
94
Step 4: Edge Threshold: The threshold method used by Canny
threshold and low threshold.
Step 5: Thinning: Using interpolation to find the pixels where the
norms of gradient are local maximum.
Figures 3.3(a) and 3.3(b) depict the before and after result of
feature extraction for a brain image using Canny Edge Detection method.
Input image Output image
(a) Brain Image (b) Edge detected Brain Image
Figure 3.3 Sample Canny Edge Detected Brain Image
Image Classification
Image classification is one of the important steps in image retrieval
process because it saves more time while searching the images from huge
volume of database. Classification is identification of different regions of the
image by which the retrieval efficiency of the system will be improved. A
commonly used classification algorithm is k-means algorithm. In this work, k-
means algorithm is used for identifying different regions of images from the
database. Performance of the retrieval process is improved by comparing the
95
d
algorithm that has been widely used in various key areas like micro array
dataset (Bernard Chen et al 2005), high dimensional data sets (Madjid
Khalilian et al 2010) and especially in image retrieval system(Hong Liu and
Xiaohong Yu 2009) . The k-means algorithm takes the input parameter k ,
and partitions a set of n objects into k clusters so that the resulting intra
cluster similarity is high but the inter cluster similarity is low. Given a set of
observations (x1, x2 n), where each observation is a d-dimensional real
vector, the k-means clustering aims to partition the n observations into k sets
(k < n) S = {S1, S2 k} so as to minimize the within-cluster sum of squares.
(3.2)
where i is the mean of points in Si.
Algorithm: For K-means clustering Algorithm
The k-means algorithm for partitioning is based on the mean value
of the objects in the cluster
Input: The number of clusters k and a database containing n
objects.
Output: A set of k clusters that minimizes the squared-error
criterion.
Method:
Step 1: Enter the number of clusters.
96
Step 2: Randomly guess k cluster Center locations.
Step 4: Each Center finds the centroid of the points it owns.
Step 5: Center now moves to the new centroid.
Step 6: Repeat until terminated.
Result Assessment
The performance was tested using reference database and the
results were analyzed and tabulated as shown in Table 3.1.
Table 3.1 Shape based CBMIR system Performance Evaluation
Query Image
Precision in %
Recall
in % Precision/Recall AVG=(P+R)/2
Time Complexity in Seconds
Image1 50.00 37.50 1.33 43.75 182
Image2 53.33 34.00 1.57 43.67 179
Image3 56.95 66.66 0.85 61.81 233
Image4 58.33 36.84 1.58 47.59 183
Image5 68.26 51.42 1.33 59.84 248
Comment on Result: Shape based CBMIR system is able to
achieve precision (accuracy) in the range of 50-70% and recall (sensitivity) in
the range of 30-70% for 5 different queries. The time complexity is in the
range of 170-250 seconds.
97
Figure 3.4 (a, b, c, d, e, and f) shows some of the sample images of
the work in various steps of Canny Edge Detection Algorithm.
CANNY EDGE DETECTION
Step 1: SMOOTHING
X Derivative:
Figure 3.4 (a) Sample Image Smoothing for X Derivative
Y Derivative:
Figure 3.4 (b) Sample Image Smoothing for Y Derivative
98
Step 2: FINDING GRADIENT
Figure 3.4 (c) Sample Image for Finding Gradient
Step 3: NON MAXIMUM SUPRESSION
Figure 3.4 (d) Sample Image for Non Maximum Suppression
Step 4: THRESHOLDING AND HYSTERESIS
Figure 3.4 (e) Sample Image for Thresholding and Hysteresis
99
Step 5: EDGE DETECTION
Figure 3.4 (f) Sample Edge Detected Image using Canny Edge Detection
Algorithm
3.1.2 Texture based CBMIR system
In this work, one of the visual features viz. texture has been
considered. The main feature of this tool is utilization of GLCM for extracting
the texture pattern of image and k-means clustering algorithm for image
classification in order to improve the retrieval efficiency.
The framework model of the texture based CBMIR system is
shown in Figure 3.5 in which the Medical images (such as x-ray, MRI Scan,
CT scan) are given as input into the system. The given input images are
segmented using the method described in (Ozden and Polat, 2005). In this
work, only the texture regions of the image are considered for feature
extraction. For each image in the image database, feature vector value has
been developed and stored in feature database. When a query image is
submitted by the user, the same texture feature extraction and feature vector
value construction process is applied to the query image to obtain the feature
vector value of the query image. For similarity comparison between the query
100
image and the database image, the Euclidean distance method is used. The
closest Euclidean distance values to the database images are ranked and
retrieved.
Texture
Texture is a natural property of surfaces and it provides visual
patterns of the image. It has repeated pixel information that contains vital
information regarding the structural arrangement of the surface (example
clouds, leaves bricks). It also gives the relationship between surface and
external environment.
Figure 3.5 Framework Model of the texture based CBMIR system
In this work, the extraction process of texture feature is performed
by computing the Gray Level Co-Occurrence Matrix (GLCM). The
graycomatrix function is used to create a GLCM. The graycomatrix function
Images Query Image
Texture Feature Extraction
Feature database
Similarity Comparison
Query Features
Retrieved Images and Ranking
101
creates a gray level comatrix by calculating how often a pixel with the
intensity (gray level) value i occurs in a specific spatial relationship to a
pixel with the value j . The spatial relationship is defined as the pixel of
interest and the pixel to its immediate right (horizontally adjacent) (Haralick
et al 1973). The outcome of GLCM for each element (I, J) is computed by
summing the pixel with the value occurred in the particular spatial
relationship to a pixel with value j in the input image (Partio et al 2002, Park
et al 2004). GLCM features are extracted using one distance d = {1} and four
After computation of Gray level co occurrence matrix, a number of
statistical texture measures based on GLCM are derived, which are suggested
by Haralick. For generating texture features, second order method has been
used that are derived from the co-occurrence probabilities. These probabilities
represent the conditional joint probabilities of all pairwise combinations of
gray levels in the spatial window of interest given two parameters: inter pixel
Eq.3.3.
Pr(x) {C | ( )}ij = , (3.3)
where, Cij is the co-occurrence probability between gray levels i
and j which is defined as Eq. 3.4
ijij
ij
PC GP
i,j=1
(3.4)
where Pij - Represents the number of occurrences of gray levels i and j -
102
G -The quantized number of gray levels
The sum in the denominator thus represents the total number of
gray level pairs (i, j) within the window.
Graycomatrix computes the GLCM from a full version of the
image. By default, if a binary image, graycomatrix scales the image to two
gray-levels. If is an intensity image, graycomatrix scales the image to eight
gray-levels. A texture is distinguished by 14 statistical measurement value
suggested by (Haralick et al. 1973). The following formulas are used to
calculate the features, which are shown in Eq. 3.5-3.8(Yin et al 2008).
2Energy P(i, j)i,j
= (3.5)
Entropy P(i, j)log(P(i, j)i,j
= - (3.6)
i j(i )( j )p(i, j)Correlationi,j i j
= (3.7)
P(i, j)Homogeneity 1 |i j|i,j = (3.8)
Algorithm: For calculating GLCM measures for each pixel
1. Read the input image.
2. Convert the data type to double and Zero pad the image
3. Extract a 3×3 window image from the input image and compute the co-occurrence texture measure
4. Estimate the texture parameters for the obtained texture image
103
5. Repeat step3 and step4 by moving the window till end of the image
6. Display various texture parameters by normalizing them
Classification: Classification is a technique to detect the dissimilar
texture regions of the image based on its features. It can be used to cluster the
feature sets of the image that is characterized as different regions. A
frequently used clustering algorithm is the k-means algorithm. In this work,
the k-means clustering algorithm is used for classifying the texture regions of
the image so that different regions of the texture image have been identified in
order to increase the performance of the retrieval by comparing the classified
texture image with u
K-means clustering: K-means clustering is a simple algorithm for
clustering the texture regions of an image. For K clusters {C1, C2 K} each
with nk patterns, it aims to find cluster centers mk to minimize the cost
function 2E k shown in Eq.3.9 and 3.10.
kk k
1m xn x C
(3.9)
(3.10)
The initial cluster midpoints are selected randomly and the
algorithm is applied repeatedly until a fixed state level is arrived.
Algorithm: For K-means clustering
1. Initialize cluster centers randomly in texture image
2. For all the pixels in the image clusters do the following
104
a) Compute the Euclidean distance of the feature vector from the cluster for every other cluster.
b) Assign the pixel to that cluster whose center yields the minimum distance from the feature vector
3. Update the cluster centers by computing the mean of the feature vectors of the pixels belonging to that cluster
4. Between two consecutive updates, if the changes in the cluster centers are less than a specified value, then stop else go to step 2
Result Assessment
The performance is tested using reference database and the results
are analyzed and tabulated as shown in Table 3.2.
Table 3.2 Texture based CBMIR system Performance Evaluation
Query Image
Precision in %
Recall
in % Precision/Recall AVG=(P+R)/2
Time Complexity in Seconds
Image1 53.33 51.42 1.04 52.38 224
Image2 55.71 45.00 1.24 50.36 187
Image3 56.00 60.00 0.93 58.00 180
Image4 57.00 56.66 1.01 56.83 210
Image5 65.71 48.94 1.34 57.33 182
Comment on Result
Texture based CBMIR system is able to achieve precision
(accuracy) in the range of 50-70% and recall (sensitivity) in the range of 40 to
60% for 5 different queries. The time complexity is in the range of 180-225
105
seconds. Compared to the shape based CBMIR system, texture based CBMIR
system seems to give better recall values and less complexity for the same
five different query images.
The following Figure 3.6.shows some of the sample images of the
work used in the model for texture feature extraction
Figure 3.6 Sample images of the work for texture feature extraction
3.2 CBMIR SYSTEM USING DUAL VISUAL FEATURE
CONTENT
In this part, the CBMIR has been implemented using dual visual
features such as the combined form of both shape and texture features.
Section 3.2.1 and 3.2.2 discusses the implementation using combined form of
both shape and texture based features.
106
Image Database
Feature Database
Select Image
Shape Feature Extraction (EHD)
Texture Feature Extraction (EBT)
Building Single Feature Vector
Query Image
Extract Texture &Shape Features
Similarity Comparison
Indexing & Retrieval
Output Ranked Image
3.2.1 CBMIR using Edge Histogram Descriptor (EHD) and Edge
Based Texture (EBT)
The primary step of this work is to extract shape using edge
histogram descriptor and texture using edge based texture in an image. In the
next step, both these image features are combined together and are considered
to build single image feature vector. This process is performed for each of the
images stored in the database to form a set of feature values. The image
selected by the user (i.e. query image) is also considered and the above
features are extracted to form a combined feature value. Further each of the
feature vector value in feature database is compared with feature value of the
query image. The most similar images are then ranked and displayed based on
their similarity factor. Figure 3.7 shows the framework model of CBMIR
system using EHD and EBT.
Figure 3.7 Framework model of CBMIR system using EHD and EBT
107
Texture feature extraction
The texture is extracted using an edge based texture feature
extraction technique. The edge based texture extraction is obtained by
defining a 3x3 matrix mask. This approach provides the necessary edge
information as well as captures the regions that do not have a clear edge
defined.
Edge based texture feature extraction
Texture features provides a key feature for retrieval systems. The
edge based texture feature extraction method gives the edge information along
with the regions. A set of six 3x3 matrix are used as edge filters. These six
filters each form a mask and are moved from pixel to pixel in an image by a
process called convolution. The formula for a two-dimensional convolution of
an image a and mask b of values n1 and n2 is given as
1 2 1 2 1 1 2 21 2
c(n ,n ) a(k ,k )b(n k ,n k )k k
(3.11)
The resulting image of the two-dimensional convolution (Eq 3.11)
of each filter mask values given in Figure 3.8 is then checked with specified
threshold value for obtaining pixel values having less than a specified
threshold value, which is to be removed. It gives a clearer texture pattern of
the image. The same process is carried out with other masks and the images
are consolidated. The different edge filter masks and sample output of the
edge based texture feature extraction are shown in Figures 3.8 and 3.9
respectively.
Figure 3.8 Different edge filter masks
108
(a) Input image (b) Texture extracted image
Figure 3.9 Edge based texture Feature Extraction
Shape feature extraction
The shape feature is extracted using edge histogram descriptor. The
edge histogram descriptor helps in capturing the spatial distribution of the
image. The edge histogram descriptor represents the local edge distribution by
subdividing the image space into 4 X 4 sub images, and represents it as
histograms.
Edge histogram descriptor
An edge histogram descriptor is performed to extract the shape of
an image. The edge histogram descriptor captures the spatial distribution of
five types of edges, namely four directional edges and one non directional
edge. In order to improve the matching performance, the edge distribution
information of the whole image space and the vertical and horizontal semi
global-edge distributions are used. It helps to retrieve the similar semantic
means of an image.The edge histogram descriptor represents the local edge
distribution by subdividing the image space into 4 X 4 sub images, and
represents it as histograms. The edge histograms are categorized as shown in
Figure 3.10 (Nandagopalan et al 2008 & Rajshree et al 2010).
109
(a) Vertical (b) Horizontal (c) 45 degree (d) 130 degree (e) non-directional
Figure 3.10 Edge histograms types
Algorithm for edge histogram descriptor is calculated as follows
1. Compute the total number of bins. Here a total of 5 x 16 = 80
histogram bins are required.
2. Find all edges in the image using the filter coefficients by
moving them each pixel by pixel. The filter coefficients for edge
detection are as shown in Figure 3.11.
Figure 3.11 Filter coefficients for edge detection
3. Place each bin corresponding to its orientation.
4. Finally normalize the histogram by dividing the value in each
bin by total number of edges.
Figure 3.12 shows sample shape feature extraction and edge
histogram for the input image.
110
(a) Input Image (b) Shape extracted image
(c) Edge Histogram
Figure 3.12 Shape Feature Extraction
In Figure 3.12 (c), edge histograms are detected till the intensity
value is null. i.e. when there is no intensity value, it means that the edges of
that paricular descriptor is completely extracted.
Combined feature vector construction
After extracting the shape and texture features, they are converted
into a common feature vector. This feature vector is constructed by combining
the texture feature values and the histogram values. The texture feature values
are averaged and consolidated into a single matrix value. The normalized
111
histogram obtained from edge histogram descriptor form the other matrix
value. Thus the texture feature values and the histogram values form a pair in
the feature vector. i.e. Feature vector = [Texture_extracted_image_value,
Edge_histogram_value]. The resulting feature is considered for similarity
comparison that provides a better image retrieval process.
Result Assessment
The performance is tested using reference database and the results
are analyzed and tabulated as shown in Table 3.3.
Table 3.3 CBMIR using Edge Histogram Descriptor and Edge Based
Texture system Performance Evaluation
Query Image
Precision in %
Recall
in % Precision/Recall AVG=(P+R)/2
Time Complexity in Seconds
Image1 87.50 84.00 1.04 85.75 355
Image2 89.00 83.33 1.07 86.17 303
Image3 93.33 88.57 1.05 90.95 335
Image4 93.44 89.47 1.04 91.46 342
Image5 91.73 90.00 1.02 90.87 300
Comment on Result
The system is able to achieve precision (accuracy) in the range of
85-95% and recall (sensitivity) in the range of 80-90% for the same set of
queries. The time complexity is in the range of 300-360 seconds. Compared
to single feature based CBMIR system, this system seems to give 50% better
112
precision and recall values but time complexity also increases by equal
amount.
Figure 3.13 shows one sample output user interface screen of this
work.
(a) Loading Screen (b) Selecting the Image from the Database
(c).Input query image (d) Output similar images
Figure 3.13 Sample user interface screen Images of CBMIR using EHD
and EBT
113
3.2.2 CBMIR using Generic Fourier Descriptor (GFD) and Gabor
Filters (GF)
This method consists of three stages: First, the input image
(Medical image) is processed for shape feature extraction using Generic
Fourier Descriptor method (GFD). In the next stage, the texture features are
extracted using Gabor Filters (GF) method. The final stage integrates the
above two features to obtain feature vector values. Then these feature vector
values are used to perform similarity comparisons between query image and
database images.
In order to perform this process, the query image is extracted based
on shape using the effective shape descriptor called Generic Fourier
Descriptor. GFD generally enhances the performance of the system by
highlighting some major advantages like retrieval accuracy, low
computational time and robust retrieval performance. The next stage is texture
feature extraction and it is extracted using the efficient Gabor filter method.
This algorithm outperforms most of other methods in the process of extracting
the texture features and it also provides better way to retrieve the images. The
retrieval process is done effectively by extracting both the shape and texture
feature of the query image and this resulting image is used to construct feature
vector value and then it is compared with the reference database to retrieve
similar images using the Euclidean distance method. Figure 3.14 shows the
framework model of CBMIR using GFD and GF.
114
Image Databas
Feature Database
Select Image
Shape Feature Extraction
(GFD)
Texture Feature Extraction (GF)
Building Single Feature Vector
Query Image
Extract Texture &Shape Features
Similarity Comparison
Output Ranked Image
Indexing & Retrieval
Figure 3.14 The framework model of CBMIR using GFD and GF
Texture Feature Extraction
Texture Feature description is one of the key features of an image
content description for image retrieval. Textures are modeled as a pattern
dominated by a narrow band of spatial frequencies and orientations. The
texture feature of an image is extracted to limit the region to be processed
using feature extraction. The texture feature extraction is carried out using a
texture analysis technique called Gabor Filters (GF). The Gabor filters are a
group of wavelets, with each wavelet capturing energy at a specific frequency
and a specific direction. Gabor filter is very useful for texture analysis
because of its tunable property of frequency and orientation. Gabor filters
have been used in many applications such as texture segmentation, target
115
detection, fractal dimension management, document analysis, edge detection,
retina identification, and image coding and image representation. (Levesque
Vincent 2000).
Texture feature extraction using Gabor Filter
Gabor filter takes the form of a Gaussian modulated complex
sinusoid in the spatial domain. A bank of Gabor filters are used to extract
local image features. Typically, an input image is convolved with a 2-D
Gabor function to obtain a Gabor feature image. Gabor filters have the ability
to model the frequency and orientation sensitivity characteristic of the human
visual system. Gabor filters have desirable properties for picture analysis and
feature extraction. They are selective in space, spatial frequency and
orientation, achieving the theoretical limit for conjoint resolution in the spatial
and spatial frequency domain. (Adams Wai-Kin Kong et al 2003, Veni et al
2010).
One Dimensional Gabor Filter
The one-dimensional Gabor Filter consists of three parts and they
are cosine, Gaussian and constant part. The cosine part is dependent on
distance and frequency. (Grigorescu Simona et al 2002), the Gaussian part is
dependent on distance and sigma, and constant part makes the two-
dimensional Gaussian interval equal to 1.0. The magnitude of the 1-D Gabor
filter output is used as a feature to detect boundaries for texture like images.
The major advantage of 1-D Gabor filter is, both the feature extraction and
edge extraction are applied along orthogonal directions. The 1-D Gabor filter
has the following form,
2
21 xf (x, , ) exp( x)
22 (3.12)
116
where x is the coordinate he standard
deviation of Gaussian envelope.
Two Dimensional Gabor Filter
The Gabor filter is a multi-scale, multi-resolution filter that has
selectivity for orientation, spectral bandwidth. Gabor function is a band-pass
filter that can be tuned to a narrow set of frequency anywhere in the frequency
domain (Anil Jain et al 2001). Gabor function is a complex sinusoid
modulated by a rotated Gaussian. It is one of the most interesting concepts
that deal with the frequency domain. This function can provide accurate time-
frequency location governed by Uncertainty principle . It states that it
satisfies the lower-most bound of time-spectrum resolution. A circular 2- D
Gabor filter in the spatial domain has the following general form
2 2
2 2
1 x yG(x, y, , u, ) exp exp 2 i ux cos uysin2 2
(3.13)
where i= 1 , u is the frequency of the sinusoidal wave, is the
orientation of the function and is the standard deviation of the Gaussian
envelope. Figure 3.15 shows the sample texture extracted output image.
(a) Input Image (b) Output Image
Figure 3.15 Gabor Filter based texture feature extraction
117
Shape Feature Extraction
The Shape is considered as one of the key features of image content
description and also an important low level image feature to retrieve relevant
images from the database. The shape feature is extracted using an effective
shape descriptor called Generic Fourier Descriptor (GFD). This shape
descriptor is obtained by applying 2-D Fourier Transform on a polar shape
image (Dengsheng Zhang et al 2002). The GFD method outperforms common
contour-based and region-based shape descriptors (Arun et al 2009). The
shape descriptor is distinguished from other descriptors by means of two
characteristics and they are stability (stable performance in different
applications) and clarity (clear physical meaning). Retrieval accuracy,
compact features, general application, low computation complexity, robust
retrieval performance and hierarchical coarse to fine representation are some of the advantages highlighted in the GFD shape descriptor.
Shape feature extraction using Generic Fourier Descriptor
The Feature extraction process is mainly performed to extract the
shape of image. In this process, initially the One Dimensional Fourier
Descriptor (1-D FD) has been applied to the query image to obtain the
knowledge of shape boundary information. Furthermore, One Dimensional
Fourier Descriptor cannot capture shape interior content which is important
for shape discrimination. To overcome the drawback of 1-D FD, Two Dimensional Fourier Descriptor (2-D FD) comes in to existence. By applying
2-D FD, first low frequency terms Fourier descriptors captures global shape
features, while higher frequency terms capture finer details of the shape
(Yang Mingqiang et al 2008 , Ekombo P. Lionel Evina et al 2009).
One Dimensional Fourier Descriptor
The One Dimensional Fourier Descriptor has been widely used in
various shape representation applications and the concept behind the 1-D FD
118
is to obtain the shape boundary of an image without considering the interior
content (Goyal Anjali et al 2010). The shape signature is the one-dimensional
function which is derived from shape boundary coordinates. It usually
captures the perceptual feature of the shape. A typical shape signature function is the centroid distance, which is given by the distance of boundary
points from the centroid (xc, yc) of the shape,
12 2 2r t x t x y t yc c
(3.14)
where N 11x x tc N t 0
, N 11y y tc N t 0 -1
Generally, 1-D FD is obtained through Fourier transform (FT) on a
shape signature function derived from shape boundary coordinates {(x (t), y
-1}
Two Dimensional Fourier Descriptor
The region-based Fourier descriptor is commonly referred to as
generic Fourier descriptor which is used for several applications (Thai V.
Hoang et al 2010). The Generic Fourier descriptor is derived by applying a
modified polar Fourier transform (MPFT) on shape image. In order to apply
MPFT, the polar shape image is treated as a normal rectangular image. The
steps are
1. The normalized image is rotated counter clockwise.
2. The pixel values along positive x-direction starting from the
image center are copied and pasted into a new matrix as row
elements.
119
3. The steps 1 and 2 are repeated until the image is rotated by
360°.
The Fourier transform is acquired by applying a discrete 2D Fourier
transform on the shape image.
i i
r i
r 2PF(R T (3.15)
w i i (R and T are
radial and angular resolutions).
frequencies selected and the number of angular frequencies selected.
Figure 3.16 shows the sample shape extracted output image.
(a) Input Image (b) Shape extracted image
Figure 3.16 Generic Fourier Descriptor based Shape feature extraction
Combined feature vector construction
After extracting the shape and texture features, they are converted
into a common feature vector. This feature vector is constructed by combining
the texture feature values and the shape descriptor values. The Gobor filter
120
based texture feature values are averaged and consolidated into a single
matrix value. The normalized values obtained from Generic Fourier descriptor
forms the other matrix value. Thus the texture feature values and the shape
feature values form a pair in the feature vector. i.e. Feature vector= [Texture _
extracted_image_value, Generic_Fourier_Descriptor_value]. The resulting
feature is considered for similarity comparison, which provides a better image
retrieval process.
Result Assessment
The performance is tested using the reference database and the
results are analyzed and tabulated as shown in Table 3.4.
Table 3.4 CBMIR using Generic Fourier Descriptor (GFD) and Gabor
Filters (GF) system Performance Evaluation
Query Image
Precision in %
Recall
in % Precision/Recall AVG=(P+R)/2
Time Complexity in Seconds
Image1 85.00 86.84 0.98 85.92 295
Image2 90.66 88.00 1.03 89.33 330
Image3 92.85 86.66 1.07 89.76 345
Image4 93.96 90.00 1.04 91.98 360
Image5 90.29 89.00 1.01 89.65 298
Comment on Result
The system is able to achieve precision (accuracy) in the range of
85-95% and recall (sensitivity) in the range of 85-90% for the same set of
queries. The time complexity is in the range of 290-360 seconds. Compared
121
to single feature based CBMIR system, the system provide 50% better
precision and recall value but the time complexity increases by equal amount.
Figure 3.17 shows sample user interface screen images of CBMIR
using GFD and GF.
(a) Input Screen . (b) Retrieval Screen
Figure 3.17 Sample user interface screen Images of CBMIR using GFD
and GF
3.3 CBMIR SYSTEM USING TRI VISUAL FEATURE
CONTENT
In this work, the CBMIR has been implemented using tri visual
features such as combined form of shape, texture and intensity of the image.
3.3.1 CBMIR using Gaussian Filter, Local Binary Pattern and
Discrete Wavelet Transform for medical images
In this work, the features extracted are: shape using Gaussian Filter,
texture using Local Binary Pattern (LBP) and Discrete Wavelet Transform
(DWT) based intensity that is used in spatial domain.
122
Images with high feature similarities to the query image may be
different from the query in terms of semantics. This work proposes a step
named image cropping that is performed only on the affected part or the
specified part of the image to tackle the semantic gap problem. Since feature
extraction is done only on the cropped part, it makes the system time efficient.
More accurate results are produced by extracting the features such as shape,
intensity and texture in spatial domain. Algorithm for this system is as follows
Algorithm
The three major tasks involved in this retrieval system are listed out
below
Step 1 : The Query image is obtained from the user.
Step 2 : Part of the image to be cropped is specified by the user.
Step 3 : The part specified is cropped.
Step 4 : Extracting the features from the cropped segment.
Step 5 : Determining image similarity and retrieval.
The framework model of CBMIR using Gaussian Filter, Local
Binary Pattern and Discrete Wavelet Transform for medical images is shown
in Figure 3.18.
123
Feature Database
Cropping Image
Intensity Feature Extraction
(DWT)
Shape Feature Extraction
(Gaussian Filter)
Building Single Feature Vector
Extract Texture, Shape &Intensity
Features
Similarity Comparison
Indexing & Retrieval
Output Ranked Image
Texture Feature Extraction
(LBP)
Image Database
Reading Input Image
Cropping Image
Query Image
Figure 3.18 The framework model of CBMIR using Gaussian Filter,
Local Binary Pattern and Discrete Wavelet Transform for
medical images
Obtaining the Input Image and Cropping
The Query image is obtained from the user. An interface is
provided to the user that allows selecting the query image from a folder and
linking it to the application. The part of query image that can be used for
feature extraction is specified by the user. This part is then cropped and given
as input to the next module (feature extraction module). Figure 3.19 (a, b, c)
illustrates this process.
124
(a) Selection of query image from the (b) Image with Pixel Information database
(c) Cropped Image
Figure 3.19 Process of cropped image
Texture Feature Extraction
The texture is a manner in which the constituent parts are united. It
is the structure or repeated patterns on an image. Texture in digital images can
be determined if the neighboring pixels satisfy a specified criterion of
similarity. Local Binary Pattern (LBP) is used in texture extraction of the
image. The LBP operator works with eight neighbors of the pixel using the
125
center pixel value as threshold and the LBP code for a neighborhood is
produced by multiplying the threshold values with weights given to the
corresponding pixels and summing up the result. (Devrim Unay and Ahmet
Ekin 2008).
Texture Feature Extraction Using Local Binary Pattern (LBP)
In this work, Local Binary Pattern algorithm is used for texture
extraction. The LBP operator defines the texture in the image represented by
thresholding the neighborhood with the gray value of its center pixel and the
results will be represented as binary code format. The pixel-to-pixel
comparison in the image produces the texture and the resulting image is in the
form of texture histogram.
Local Binary Pattern Algorithm for Texture Extraction
STEP 1 : For each pixel in the cell, compare it with other 8 neighboring pixels.
STEP 2 : Follow the pixel along a circle i.e. clockwise or anti-clockwise.
STEP 3 : write 1 otherwise write 0 .This gives an 8 bit binary number.
STEP 4 : Compute the histogram over the cell of frequency of each number ccurring.
STEP 5 : Optionally normalize the histogram.
STEP 6 : Concatenate normalized histogram of all cell.
STEP7 : This gives the feature vector for the window, which can be used for classification.
126
Figure 3.20 shows a texture feature extracted image.
Figure 3.20 LBP based Texture Feature Extracted Image
Shape Feature Extraction
The shape of a set of points is all the geometrical information that is
invariant to size changes. The shape does not depend on the size of object and
on changes in orientation/direction. However, a mirror image could be called
a different shape. Shapes may change if the object is scaled non-uniformly.
Shape Feature Extraction Using Gaussian Filter
In this work, Gaussian filter based shape feature extraction is
implemented. Gaussian filter is an effective way of removing the noise. Since
the weights gives higher significance for the pixels, it reduces edge blurring.
x is the distance from the origin
in the horizontal axis, y is the distance from the origin in the vertical axis, and
x and y values are generated randomly.
2 2x y( )2G(x,y) e (3.16)
127
The generated mask is convoluted with the original image.
a h*mask (3.17)
where h is the original image. Figure 3.21 shows a shape feature
extracted image.
Figure 3.21 Gaussian Filter based Shape Feature Extracted Image
Intensity Feature Extraction
The intensity is the amount of light the pixel reproduces (how
bright it is). Gray scale images also known as black and white images are
composed exclusively of shades of gray, varying from black at the weakest
intensity to white at the strongest. The binary representations assume that 0 is
black and the maximum value (255 at 8 bpp, 65,535 at 16 bpp, etc.) is white.
Intensity Feature Extraction Using Discrete Wavelet Transform
In this work, the Discrete Wavelet Transform is used to extract the
intensity feature in spatial domain. A discrete wavelet transform (DWT) is
any wavelet transform for which the wavelets are discretely sampled. A key
128
advantage is that it captures both frequency and location information (location
in time) (Hasan Demirel and Gholamreza Anbarjafari 2011).
Discrete Wavelet Transform Algorithm
STEP 1 : Separate the image pixel positions into odd and even columns.
STEP 2 : Add the odd and even column values to construct a low pass filter.
STEP 3 : Subtract the odd and even column values to construct a high pass filter.
STEP 4 : Separate the low pass filter into odd and even rows.
STEP 5 : Similarly the high pass filter is separated into odd and even rows.
STEP 6 : Get the four results such as LL, LH, HL, and HH.
STEP 7 : Use the obtained results for further feature extraction.
Figure 3.22 shows intensity feature extracted image.
Figure 3.22 DWT based Intensity feature extracted image
129
Combined feature vector construction
After extracting all the three features such as shape, texture, and
intensity, these features are integrated in order to build a single feature vector
using fusion method. The resulting feature is considered for similarity
comparison that provides a better image retrieval process.
Result Assessment
The performance is tested using reference database and the results
are analyzed and tabulated as shown in Table 3.5.
Table 3.5 CBMIR using Gaussian Filter, Local Binary Pattern and
Discrete Wavelet Transform for medical images system
Performance Evaluation
Query Image
Precision in %
Recall
in % Precision/Recall AVG=(P+R)/2
Time Complexity in
Seconds
Image1 90.93 85.57 1.06 88.25 405
Image2 91.58 85.00 1.08 88.29 464
Image3 92.00 90.10 1.02 91.05 397
Image4 93.73 86.66 1.08 90.20 490
Image5 94.65 87.00 1.09 90.83 465
Comment on Result
This tri feature extraction system is able to achieve precision
(accuracy) in the range of 90-95% and recall (sensitivity) in the range of
85-90% for the same set of queries. The time complexity is in the range of
400-490 seconds. Compared to the dual feature based CBMIR system, this
130
system seems to give about 5% better precision and recall values but time
complexity is also increased by 40%.
Figure 3.23 shows one sample snapshot of the images retrieved in
the work.
Figure 3.23 Sample snapshot of retrieved images of CBMIR using
Gaussian Filter, Local Binary Pattern and Discrete Wavelet
Transform for medical images system
3.4 ANALYSIS AND COMPARISON OF SINGLE, DUAL AND
TRI FEATURE EXTRACTION METHODS.
Table 3.6 shows the overall result assessment comparison of single,
dual and tri feature extraction methods implemented for the same reference
database under similar experimental setup.
1
Table 3.6 Overall result assessment comparison
Query Image
METHOD-I
SINGLE FEATURE EXTRACTION
METHOD-II
DUAL FEATURE EXTRACTION
METHOD-III
TRI FEATURE EXTRACTION
Shape(CED) Texture(GLCM) Shape & Texture
(EH D + EBT) Shape & Texture
(GFD + GF)
Shape, Texture & Intensity
(Gaussian Filter +
L B P + D W T)
P R T.C P R T.C P R T.C P R T.C P R T.C
Imag1 50.00 37.50 182 53.33 51.42 224 87.50 84.00 355 85.00 86.84 295 90.93 85.57 405
Imag2 53.33 34.00 179 55.71 45.00 187 89.00 83.33 303 90.66 88.00 330 91.58 85.00 464
Imag3 56.95 66.66 233 56.00 60.00 180 93.33 88.57 335 92.85 86.66 345 92.00 90.10 397
Imag4 58.33 36.84 183 57.00 56.66 210 93.44 89.47 342 93.96 90.00 360 93.73 86.66 490
Imag5 68.26 51.42 248 65.71 48.94 182 91.73 90.00 300 90.29 89.00 298 94.65 87.00 465
Average 57.38 45.28 205 57.55 52.40 197 91.00 87.07 327 90.55 88.10 326 92.57 86.86 444
*. P-Precision; R-Recall; T.C-Time Complexity in Seconds.
131
132
It is inferred from the Table 3.6, that the tri feature extraction
method shows the best result in terms of both precision (accuracy) and recall
(sensitivity) parameters when compared to other methods such as single
feature extraction method and dual feature extraction method. One constraint
noted is the increased time complexity.
Conclusion
This chapter presents implementation and result analysis of the best
methods selected out of many methods that have been discussed in the
literature review. To compare the throughput of the selected schemes, similar
experimental setup has been ensured by implementing all the schemes on the
same machine with identical test conditions. Precision and recall were
calculated using the same reference database having 560 medical images that
includes real-time medical images collected from local hospital and medical
images collected from CasImage database and IRMA database repositories.
Considering the similarity in execution and presentation concepts used, they
are grouped into three categories viz. single, dual and tri feature extraction.
They are compared in order to assess the exact retrieval performance of these
works and derive the best out of them for further investigation. It is concluded
that the tri feature extraction method shows the best result in terms of both
precision and recall parameters, when compared to other methods such as
single feature extraction method and dual feature extraction method.